Boom or gloom? Examining the Dutch disease in a two

Centre for Applied Macro - and petroleum economics (CAMp)
CAMP Working Paper Series
No 6/2013
Boom or gloom? Examining the Dutch
disease in a two-speed economy
Hilde C. Bjørnland and Leif Anders Thorsrud
© Authors 2013.
This paper can be downloaded without charge from the CAMP website http://www.bi.no/camp
Boom or gloom? Examining the Dutch disease in a
two-speed economy∗
Hilde C. Bjørnland†
Leif Anders Thorsrud‡
August 13, 2013
Abstract
Traditional studies of the Dutch disease do not typically account for productivity spillovers between the booming energy sector and non-oil sectors. This study
identifies and quantifies these spillovers using a Bayesian Dynamic Factor Model
(BDFM). The model allows for resource movements and spending effects through
a large panel of variables at the sectoral level, while also identifying disturbances
to the real oil price, global demand and non-oil activity. Using Norway as a representative case study, we find that a booming energy sector has substantial spillover
effects on the non-oil sectors. Furthermore, windfall gains due to changes in the
real oil price also stimulates the economy, but primarily if the oil price increase is
caused by global demand. Oil price increases due to, say, supply disruptions, while
stimulating activity in the technologically intense service sectors and boosting government spending, have small spillover effects on the rest of the economy, primarily
because of reduced cost competitiveness. Yet, there is no evidence of Dutch disease.
Instead, we find evidence of a two-speed economy, with non-tradables growing at
a much faster pace than tradables. Our results suggest that traditional Dutch disease models with a fixed capital stock and exogenous labor supply do not provide
a convincing explanation for how petroleum wealth affects a resource rich economy
when there are productivity spillovers between sectors.
JEL-codes: C32, E32, F41, Q33
Keywords: Resource boom, oil prices, Dutch disease, learning by doing, two-speed economy, Bayesian Dynamic Factor Model (BDFM)
∗
This paper is part of the research activities at the Centre for Applied Macro and Petroleum economics
(CAMP) at the BI Norwegian Business School. The authors would like to thank Francesco Ravazzolo,
Ragnar Torvik and Benjamin Wong for valuable comments. The usual disclaimer applies. The views
expressed in this paper are those of the authors and do not necessarily reflect the views of Norges Bank.
†
BI Norwegian Business School and Norges Bank. Email: [email protected]
‡
BI Norwegian Business School and Norges Bank. Email: [email protected]
1
1
Introduction
Over the last decade, the value of the Norwegian oil and gas industry - including services
- grew by approximately 90 percent, while employment in this industry grew by 70 per
cent. No other industry exhibited such growth rates.
The oil and gas boom in the North Sea has been the principal, but by no means,
only cause of this substantial growth. Strong rises in oil and gas prices have caused
Norway’s terms of trade to double since 2001. These price rises have profound effects
on the economy, as they constitute both a large shift in relative prices, which induces
resource movements between industries, and a large increase in real incomes, which boosts
aggregate demand in the overall economy.
While the recent financial crisis has suggested that energy rich countries - such as Norway - have occupied a different and better position than many other indebted industrial
countries,1 it is not clear that the gains from the energy sector benefited domestic sectors
equally. For instance, employment in the construction and business sectors in Norway has
increased by 30-40 percent over the last decade, while employment in the manufacturing
industry and the retail, hotel and service industry has either fallen or hardly grown.
The energy boom has prompted much discussion of Norway having become a twospeed economy. There are concerns that the gains from the boom largely accrue to the
profitable sectors servicing the energy industry, such as the business services, financial and
construction sectors, while the rest of the country is being negatively affected by increased
wage costs, an appreciated exchange rate and a lack of competitiveness as a result of the
boom. Such a phenomenon has commonly been referred to in the literature as the Dutch
disease, based on similar experiences in the Netherlands in the 1960s.2 Concerns are also
raised in other resource rich countries recently, such as the petroleum producer Canada
and the mineral rich Australia.3
Much theoretical work has analyzed the benefits and costs of energy discoveries (see,
e.g., Corden (1984) for a survey), but there have been relatively few empirical studies.
Those that have investigated the empirical relationship between a booming energy sector and the macro economy have typically employed a structural vector autoregression
(SVAR), which only includes a single sector such as manufacturing in each model, see,
e.g., Hutchison (1994) and Bjørnland (1998), or a panel data approach that studies common movements in manufacturing across numerous countries, see, e.g., Ismail (2010).
However, neither of these approaches accounts for all of the cross-sectional co-movement
of variables within a country. That is, spillovers between sectors of the economy can be
substantial due to intermediate inputs between the sectors and induced effects through
increased demand and income in the energy sector or the sectors that are indirectly af1
Mehlum et al. (2006) argue that in countries with strong property rights protection and little corruption,
natural resources may have contributed to growth.
2
Following the discovery and development of natural gas industries in the 1960s, the Netherlands experienced a period of real exchange rate appreciation relative to other nations and a corresponding loss of
competitiveness for traditional industries that eventually contracted.
3
See e.g., Lama and Medina (2012) for a discussion of the usefulness of exchange rate stabilization in
relation to Dutch disease in Canada, and Corden (2012) for a discussion of the fast growing Australian
mining sector on the one side and the lagging manufacturing sectors on the other.
2
fected. In addition, there may be shared productivity dynamics. For Norway, where oil
extraction is conducted offshore and with greater technical difficulties than for typical onshore extraction, productivity (knowhow) spillovers through high-tech industries might be
substantial. Lastly, there are other sources of shocks that could be causing the economic
boom that need to be controlled for, such as common global demand shocks.
We contribute to this area of the literature by explicitly identifying and quantifying
the linkages between a booming energy sector and sectoral performance in the rest of
the economy, while also allowing for independent disturbances to the real price of oil,
world activity and domestic (non-oil) activity. Our main focus is to test the hypothesis of
Dutch disease by separately examining the windfall gains associated with energy booms
and real oil price changes for various sectors, while also controlling for changes in global
and domestic activity. Having established the linkages, we analyze how the domestic
economy responded to the energy boom and energy price changes in different periods.
To explore these questions, we estimate a Bayesian Dynamic Factor Model (BDFM),
that includes separate activity factors for oil and non-oil sectors in addition to global
activity and the real price of oil. The BDFM is particularly useful to answer the research
questions we address.4 First, the interdependence between the different branches of an
economy - traditionally measured by the input-output tables from the National Accounts do not account for the indirect spillover effects (productivity or demand) between different
sectors. Thus, co-movement across sectors due to common factors, i.e., oil or non-oil, is not
captured by observable variables alone. Conversely, in the BDFM, latent common factors
can be identified and estimated simultaneously with the rest of the model’s parameters.
Thus, the size and sign of spillover effects can be derived and analyzed. Second, to
quantify the spillover effects across a large cross section of sectors and variables, standard
multivariate time series techniques are inappropriate due to the curse of dimensionality.
The BDFM is designed for data rich environments such as ours. Third, macroeconomic
data are often measured with noise and errors. In the factor model framework, we can
separate these idiosyncratic noise components from the underlying economic signal.
We extend the literature in three ways. First, to the best of our knowledge, this is
the first paper to explicitly analyze and quantify the linkages between a booming energy
sector and sectoral performance in the domestic economy using a structural model, while
also allowing for explicit disturbances in real oil prices, world activity and activity in the
non-oil sector. Thus far, very little is known about the effect that energy booms have on
the rest of the economy in a resource rich economy, and equally important, if it is the
booms themselves or the windfall gains associated with real oil price changes that are
the most important. Second, given the large number of variables and industries included
in the analysis, this is also the most comprehensive analysis to date of the relationship
between energy booms and macroeconomic activity at the industry level in a resource rich
economy. We lastly show that standard multivariate methods do not adequately quantify
resource booms in a resource rich country such as Norway. The BDFM does, and the use
of this modeling framework to analyze the Dutch disease is novel in the literature.
4
As discussed in, e.g., Boivin and Giannoni (2006), there is a close resemblance between theoretical DSGE
models and Dynamic Factor Models. Moreover, Bai and Wang (2012) discuss how the DFM can be
related to the Structural VAR literature.
3
Our main conclusion emphasizes that a booming energy sector has significant and large
productivity spillovers on non-oil sectors, effects that have not been captured in previous
analysis. In particular, we find that the energy sector stimulates investment, value added,
employment and wages in most tradable and non-tradable sectors. The most positively
affected sectors are construction, business services and real estate.
Furthermore, windfall gains due to changes in the real oil price also stimulate the
economy, particularly if the oil price increase is associated with a boom in global demand. Oil price increases due to, say, supply disruptions, while stimulating activity in
the technologically intense service sectors and boosting government spending, have small
spillovers effects to the rest of the economy, in part because of substantial real exchange
rate appreciation and reduced cost competitiveness. Yet, there is no evidence of Dutch
disease as experienced in the Netherlands in the 1970s, where the manufacturing sector
contracted. Instead, we find evidence of a two speed economy, with employment in the
manufacturing sector lagging behind the booming service sectors.
Our results suggest that traditional Dutch disease models with a fixed capital stock
and exogenous labor supply do not provide a convincing explanation for how petroleum
wealth affects a resource rich economy when there are productivity spillovers between the
various sectors.
The remainder of the paper is structured as follows. In Section 2, we briefly discuss the
theoretical literature on Dutch disease and present some stylized facts. Section 3 and 4
describe the data and the model, the identification strategy and the estimation procedure
in detail. Our main results are reported in Section 5, while in Section 6, we show that
these results are robust to numerous specification tests. Section 7 concludes.
2
Macroeconomic impacts of an energy discovery
There is a substantial theoretical literature on the Dutch disease, see, for instance Bruno
and Sachs (1982), Corden and Neary (1982), Eastwood and Venables (1982), Corden
(1984), Van Wijnbergen (1984) and Neary and van Wijnbergen (1984). The general
finding in most of these papers is that there is an inverse long run relationship between
increased exploitation of natural resources and growth in the manufacturing sector, similar
to what the Netherlands experienced in the 1960s.
Although the disease most often refers to the consequences of the discovery of natural
resources, it can also refer to any development that results in a large inflow of foreign
currency, such as a sharp increase in commodity prices. As such, the analysis of the
effects of a commodity price shock on a resource rich economy is simply a special case of
the Dutch disease.
The standard theory model that these papers build on assumes a non-traded goods
and service sector and two traded goods sectors: the booming sector and the lagging
sector, also called the non-booming tradable sector. The booming sector is usually the
extraction of oil or natural gas, but can also be mining. The lagging sector generally
refers to manufacturing, but can also be agriculture when traded. The non-traded goods
and service sector includes the government sector and other non-traded sectors.
4
The direct impact of oil and gas resources (or any other sectoral boom) is experienced
through an increased demand for resources and goods and services in the energy producing
sector. This is usually referred to as a the Resource Movement Effect. The increased
demand for goods and services by the energy sector will lead to an indirect (secondary)
effect of increased demand for resources by the sectors that will produce goods and services
for the energy sector. If income in the energy sector has increased, there will also be a
further (induced) effect of increased demand for goods and services. These induced effects
are usually described as the Spending Effects, and will cause a real appreciation that will
hurt some sectors and benefit others.
More formally, Corden and Neary (1982) assume that the booming sector (B) and
the tradeable sectors (T) produce tradeables given world prices, whereas the prices for
non-tradables (N) are given by domestic factors. The energy boom is understood as an
exogenous (unpredicted) technical improvement in B. The resource movement effect will
increase demand for labor in B, as the marginal product of labor increases due to the
boom, given constant wages in terms of the tradables. Thus, there will be a movement
of labor out of T and N into B. The movement of labor from T to B will directly reduce
output in T, whereas the movement of labor from N to B at constant prices will initially
reduce the supply of N and create an excess demand for N. In response to this excess
demand, the price for non-tradables in terms of tradables will rise, which will produce
real appreciation and further movements of resources out of T into N.
The aggregate income of the factors initially employed in the booming sectors will also
rise. This will lead to a spending effect, directly by the factor owners in B or indirectly by
the government that collects (part of) the income through taxes. With positive income
elasticity of demand for N, the price of N relative to the price of T must rise, yielding a
further real appreciation. Given full employment of all resources, this real appreciation
will induce additional movement of labor from T to N.
Although the simple model of Dutch disease predicts that manufacturing will eventually contract as the energy sector expands, there are several ways that the core model
may be altered. By changing some of the underlying assumptions (for instance, by allowing the factors of production to be mobile), the predicted effects of energy booms on
the manufacturing sector may be less severe, and in fact, in some cases there may not
be Dutch disease at all. In particular, if one is initially in a situation where domestic
resources are not fully employed prior to the energy boom, the boom may actually have
a stimulative effect on industry.
Output in the manufacturing industry may also increase if one assumes that the energy
sector has its own specific factor, labor is mobile between the three sectors but capital
is only mobile between the non-tradable and the tradable sector. This constitutes a
miniature Heckscher-Ohlin economy, where one sector will be labor intensive while the
other will be capital intensive. In this case, the resource movement effect will cause the
output of the capital intensive industry to expand (as labor is moving out of the labor
intensive industry and into the booming energy sector during the boom). If the tradable
sector is the capital intensive industry, and the (negative) spending effect on output in
the tradable sector is smaller than the resource movement effects, output in the tradable
sector may actually increase, see Corden (1984) for a further discussion.
5
More recently, Torvik (2001) advanced a model in which there is learning by doing
(LBD) in both the traded and non-traded sectors, as well as learning spillovers between
the sectors. Under certain conditions, this will imply a real exchange rate depreciation in
the long run, due to a shift in the steady state relative productivity between the traded and
non-traded sectors. In contrast to the standard models of the Dutch disease, production
and productivity in both sectors can then increase.5
Thus, while the traditional theory of Dutch disease implies that the tradable sector will
eventually contract as the energy sector expands, there are several ways the dynamics of
the core model may change such that the predicted effects of energy booms on the tradable
sector may be less severe than in the basic case, and in some cases there may be no Dutch
disease at all.
2.1
Dutch disease and stylized facts of Norway
Figure 1 depicts the evolution of the important variables involved in the debate on Dutch
disease. Key to the discussion is the real oil price and the real exchange rate, depicted
for the period 1983-2012 in Figures 1a and 1b, respectively. Two features stand out.
The real exchange rate depreciated considerably between the beginning of the 1980s and
2000, after which it appreciated sharply.6 Taking everything else as given, the prolonged
period of real exchange rate depreciation in the first half of the sample fits nicely into
the framework of a model that allows for productivity advances due to learning by doing
within and between sectors, such as in Torvik (2001), discussed above. The timing of the
strong appreciation in the latter half of the sample corresponds to the increase in the real
oil price, and thus indicating a more classical Dutch disease pattern.
Figure 1c shows the evolution of employment by industry since 1996 (from which
data are available). The figure suggests a two speed economy, with resources rapidly
moving into both the booming oil and gas industry and the profitable service sectors,
while employment in other sectors such as manufacturing is lagging behind.
Lastly, Figure 1d illustrates the importance of investments in the energy sector over
the business cycle for GDP in Mainland Norway (value added of total GDP minus the oil
and gas sector). The figure clearly shows a leading and pro-cyclical relationship between
investment in the oil sector and GDP in Mainland Norway (the correlation coefficient
is 0.6 when oil investment leads the business cycle by 4 quarters), except during the
Norwegian banking crisis in the early 1990s, when other factors were at play. However,
the figure also indicates that since 2003/2004, the dynamics of the economy are not all
driven by oil. While oil investment is still pro-cyclical, the stimulus from the oil sector
seems small compared to the stimulus during the booms in the early 1980s and mid 1990s.
Other factors will have to explain the boom in the mainland economy in this period.
Thus, there is evidence that the energy sector has positive spillovers to the mainland
economy, albeit possibly to a smaller extent in the most recent boom and bust. However,
5
Traditional LBD models such as Van Wijnbergen (1984), which accounts for LBD by assuming that
productivity in the tradable sector depends on production in the first period alone, or Sachs and Warner
(1995), which employs an endogenous growth model, find unambiguously that productivity will decline.
6
This is the effective exchange rate, where an increase implies appreciation.
6
Figure 1. Stylized facts
(a) Real exchange rate
(b) Real price of oil
104
140
Real exchange rate
Real price of oil
102
120
100
100
98
96
80
94
60
92
40
90
20
88
86
1983.01
1990.03
1998.01
2005.03
0
1983.01
2012.04
(c) Employment
1990.03
1998.01
2005.03
2012.04
(d) Output gap and oil investment
0.06
Output gap mainland
Oil investment
0.04
0.02
0
−0.02
−0.04
−0.06
1983.01
1990.03
1998.01
2005.03
2012.04
Note: All employment series are on a log scale, and normalized to 100 in 1996:Q1. Figure 1d displays the
smoothed Hodrick-Prescott filtered output-gap in GDP Mainland Norway as well as the smoothed fraction
between cyclically adjusted oil investments and the trend growth in GDP Mainland Norway.
three concurrent evolutions after 2001, the appreciation of the currency, the strong rise in
commodity prices and strong growth in the oil sector relative to the manufacturing sector,
suggest a typical case of Dutch disease, where some sectors are growing at the expense of
others. We examine this subject below.
3
Theory meets data
How can one apply the theoretical model to the data? The approach we adopt relies
on the standard model presented in Corden and Neary (1982), but augmented in some
dimensions by allowing for productivity spillovers between sectors of the economy. In
particular, we develop a framework where the energy sector uses its own factor of production and develops its own specific productivity dynamics, but there may be instantaneous
spillovers to all the other domestic sectors. Thus, developments in the energy sector
will be exogenous at time t, but after a period, it may respond to the other sectors of the
economy. For instance, capacity constraint in the domestic economy could eventually also
7
affect the energy sector. Furthermore, we assume that the tradable and the non-tradable
sectors of the economy have their own factors of production and develop their own productivity dynamics, but there may be instantaneous spillovers between the tradable and
non-tradable sectors (in addition to the spillovers from the energy sector). Thus, we allow
for learning by doing in both the traded and the nontraded sector and learning spillovers
between these sectors, as suggested in Torvik (2001). Finally, we will allow for common
shocks to the global oil market.
Given the framework described above, we can identify four factors with associated
shocks that have the potential to affect all sectors: Two shocks will relate to the dynamics in domestic economy. The energy boom (or oil activity shock) 7 and the non-oil activity
shock. We let energy booms represent an unexpected technical improvement or windfall
discovery of new resources in the energy sector, while the non-oil activity shock controls
for the remaining domestic impulses (tradable and non-tradable) contemporaneously unrelated to the oil sector. In addition, we allow for two shocks that relate to the dynamics
specific to the global oil market, an oil specific shock and a global demand shock. The oil
specific shock allows for a windfall gain due to higher real oil prices from, say, a supply
disruption in oil production, while the global demand shock allows for higher oil prices
due to increased global activity.
A central premise of the theory is that the energy sector supports many more jobs
than it generates, directly owing to its long supply chains and spending by employees and
suppliers. Thus, to accommodate resource movement and spending effects, we employ
a broad range of sectoral employment, production, wage and investment series for the
Norwegian economy. The intuition is as follows: First, energy extraction may stimulate
value-added among downstream industries, such as refining, or industries that provide
the energy sector with goods and services. This will generate additional jobs in excess of
those directly produced in the energy sector. Furthermore, energy extraction can induce a
reallocation of labor from the less profitable sectors into the booming sectors. We capture
these effects by including data for value added and employment at the industry level.
Second, there will be induced spending effects through the wages paid to workers in
the energy sector or the sectors that are indirectly affected. Moreover, as the booming
sector also pays significant taxes on its increased income, these benefits will easily spread
to the whole economy. However, as Norway has a centralized wage bargaining system,
we do not include wage data for all sectors, which would be highly correlated. Instead,
for wages, we separate between the booming sector (oil and gas), the mainland (non-oil)
sector and the public sector. Note that the public sector is included to also account for
the pass through of changes in oil income to the economy.
Third, specific sectors of the economy may benefit due to productivity spillovers when
the patterns of domestic demand shifts in their favor. The loser are those producers that
do not benefit from these spillovers, what Corden (2012) terms the lagging sector. To
account for these productivity spillovers, we also include investment at the sectoral level.
We separate investments in the same way as wages.
Naturally, we include the real price of oil and the real exchange rate, which are core
7
We will use the terms energy booms and oil activity shocks interchangeably
8
factors in the Dutch Disease literature. The real price of oil is constructed based on Crude
Oil-Brent prices, deflated using the US CPI. As such, it is meant to reflect the global real
price of oil. The notion is that an increase in the real oil price will directly cause the
exchange rate to appreciate via the terms of trade. This will have adverse effects on the
tradable sector, leading to a period of de-industrialization. While this is only one part
of the question we analyze, many papers have only focused on the effects of an oil price
increase when analyzing the Dutch disease, see, e.g., Charnavoki and Dolado (2012) and
the references therein.
The de-industrialization effect described above could be a feature of Dutch disease,
but it could also be a common feature of many open economies. To control for the state of
the international business cycle, we also include a measure of global activity. We measure
global or world activity as the simple mean of four-quarter logarithmic changes in real
GDP in: China, Denmark, Germany, Japan, the Netherlands, Sweden, the UK and the
US. This set of countries includes Norway’s most important trading partners and the
largest economies in the world.
In sum, this gives a panel of 50 international and domestic data series, covering a
sample period from 1996:Q1 to 2012:Q4.
Our focus is on quantifying economic fluctuations over the horizons relevant for medium
term macroeconomic policy and over business cycle horizons. To capture the economic
fluctuations of interest, we transform all variables to four-quarter logarithmic changes;
log(xi,t ) − log(xi,t−4 )).8 Lastly, all variables are demeaned before estimation. Further
details on the data are provided in Appendix A.
3.1
Quantifying the resource boom - a simple attempt
The petroleum sector’s share of total GDP in Norway has fluctuated around 20 percent
the last decade. However, although the sector is capital intensive, it does not operate
in isolation. According to Eika et al. (2010), the total use of (non-oil) resources in the
petroleum sector was equivalent to 17 percent of the GDP of Mainland Norway (based
on input-output tables from 2008).9 However, this measure of petroleum dependency
likely represents a lower bound on the Norwegian economy’s oil dependence. Typically, it
will underestimate the links across sectors, as it does not account for the effects induced
over time from increased demand and income in the energy sector or the sectors that are
indirectly affected (e.g., the government sector).
To obtain an initial impression of the oil dependence of the Norwegian economy, one
can run a series of simple structural vector autoregressions (VARs) relating the oil sector
to the mainland economy. The analysis below is an attempt in that direction, although
as we will see, it is far from adequate in capturing the spillovers we seek.
Panels (a)-(c) of Figure 2 report the responses of GDP in Mainland Norway to three
different shocks: Global activity, oil price (specific) and oil activity, respectively. Panels
8
We experimented with specifying the model using data transformed to quarterly changes, i.e., log(xi,t ) −
log(xi,t−1 )). However, for Norwegian data, such transformations yield a very weak factor structure,
making the dynamic factor model, see Section 4, less appropriate.
9
This number is calculated based on the intermediate inputs to the petroleum sector, adjusted for the
indirect use of resources between the different sectors.
9
Figure 2. VAR (non) evidence
Impulse responses - GDP Mainland Norway
(a) Global activity
(b) Oil price
(c) Oil activity
0.35
2
0.15
0.3
0.1
0.25
1
4−VAR
3−VAR
2−VAR
4−VAR
3−VAR
4−VAR
1.5
0.2
0.05
0.15
0.5
0.1
0
0
0.05
0
−0.05
−0.5
−0.05
−1
0
5
10
15
20
−0.1
0
5
10
15
20
−0.1
0
5
10
15
20
Variance decompositions - GDP Mainland Norway
(d) Global activity
(e) Oil price
0.7
(f ) Oil activity
0.2
0.07
4−VAR
4−VAR
3−VAR
0.18
0.65
4−VAR
3−VAR
2−VAR
0.06
0.16
0.6
0.05
0.14
0.55
0.12
0.5
0.04
0.1
0.03
0.08
0.45
0.06
0.02
0.4
0.04
0.35
0
0.01
0.02
5
10
15
20
0
0
5
10
15
20
0
0
5
10
15
20
Note: The figures report impulse responses and variance decompositions of GDP in Mainland Norway to
three structural shocks: An international activity shock, an oil price shock, and an activity shock to the
petroleum sector. Three different VAR specifications are estimated: 4-VAR (world activity, real price of
oil, oil activity, mainland activity), 3-VAR (real price of oil, oil activity, mainland activity), 2-VAR (oil
activity, mainland activity). All variables are transformed to log year on year changes, and all VARs are
specified with eight lags. The structural shocks are identified employing a recursive ordering.
(d)-(f) present the variance decomposition of the same three shocks. Three different VAR
models are specified. In the 2-VAR, we jointly model oil activity and mainland activity, in
the 3-VAR we add the real price of oil, while in the 4-VAR world activity is also included,
see Figure 2 for more details. None of the VAR specifications yield results that provide an
economic meaningful depiction for quantifying a resource boom in a two-speed economy.
That is, an unexpected positive innovation in oil activity increases GDP in Mainland
Norway in all VAR specifications (Panel c), but the shock explains a negligible share
of the variance in the GDP of Mainland Norway (3-6 percent, see Panel f). This is at
odds with conventional wisdom, earlier research (see, e.g., Bjørnland (1998) and Larsen
(2006)), and most important, the National Account statistics described above.
However, the positive and large effects of a world activity shock (Panels a and d)
is in accordance with new and existing evidence of international business cycle synchronization, see, e.g., Kose et al. (2003), Stock and Watson (2005b), and Thorsrud (2013).
Furthermore, an unexpected increase in the real price of oil increases mainland activity, but primarily in the 3-VAR specification, see Panel (b). However, as shown in, e.g.,
10
Aastveit et al. (2012), a large fraction of the variation in the real price of oil can be
attributed to global activity. Only the 4-VAR specification takes this into account by
allowing the oil price to also respond to global activity. Thus, the oil price shock in the
3-VAR model is likely a combination of world activity innovations and pure unexpected
oil price innovations. This renders the structural interpretation of this model dubious and
suggests that the 4-VAR specification is more appropriate.10
Why do the structural VAR models fail to explain the resource boom in a two speed
economy? The answer is simple. They do not take all the cross-sectional co-movement of
main sectoral variables into account. That is, oil activity alone does not accurately measure the resource moving and spending effects induced by an oil boom, or any potentially
shared productivity developments.
The Dynamic Factor Model (DFM) proposed in this study solves these issues. Within
the DFM framework, the co-movement of a large cross section of variables is assumed to be
driven by a few latent (or observable) factors. The factors and the unexpected innovations
(shocks) to the factors can be identified, and structural analysis can be conducted. Geweke
(1977) is an early example of the use of the DFM in the economic literature. Kose et al.
(2003) and Mumtaz et al. (2011) are more recent examples, while Stock and Watson
(2005a) provide a brief overview of the use of this type of models in economics. In the
next section, we provide a more detailed description of the DFM, and identification and
estimation within this framework, before turning to the results in section 5.
4
The Dynamic Factor Model
We specify a Dynamic Factor Model (DFM). As noted above, this model is particularly
useful in a data rich environment such as ours, where common latent factors and shocks
are assumed to drive the co-movements between economic variables in the Norwegian
economy.
The DFM is given by equations 1 and 2:
yt = λ0 ft + · · · + λs ft−s + t
(1)
where the N × 1 vector yt represents the observables at time t. λj is a N × q matrix with
dynamic factor loadings for j = 0, 1, · · · , s, and s denotes the number of lags used for
the dynamic factors ft . In our application the q × 1 vector ft contains both latent and
observable factors. Lastly, t is an N × 1 vector of idiosyncratic errors.
The dynamic factors follow a VAR(h) process:
ft = φ1 ft−1 + · · · + φh ft−h + ut
(2)
where ut is a q × 1 vector of VAR(h) residuals. The idiosyncratic and VAR(h) residuals
are assumed to be independent:
!
t
0
R 0
∼ i.i.d.N
,
(3)
ut
0
0 Q
10
The variables included in the VARs are noise measures of the underlying business cycles. However, the
results reported in Figure 2 are robust to using HP-filtered data.
11
Further, in our application R is assumed to be diagonal.
The model described above can easily be extended to the case with serially correlated
idiosyncratic errors. In particular, we consider the case where t,i , for i = 1, · · · , N , follows
independent AR(l) processes:
t,i = ρ1,i t−1,i + · · · + ρl,i t−l,i + ωt,i
(4)
where l denotes the number of lags, and ωt,i is the AR(l) residuals with ωt,i ∼ i.i.d.N (0, σi2 ).
I.e.:


σ12 0 · · · 0


 0 σ22 . . . 0 

R=
(5)
. ,
 .. . . . .
.
. .. 
.
2
0 · · · · · · σN
4.1
Identification
Equations 1 and 2 are not identified without restrictions. To separately identify the factors
and the loadings, and to be able to provide an economic interpretation of the factors, we
enforce the following identification restrictions on equation 1:
λ0,1
(6)
λ0 =
λ0,2
where λ0,1 is a q × q identity matrix, and λ0,2 is left unrestricted. As shown in Bai and Ng
(2010) and Bai and Wang (2012), these restrictions uniquely identify the dynamic factors
and the loadings but leave the VAR(h) dynamics for the factors completely unrestricted.
Accordingly, the innovations to the factors, ut , can be linked to structural shocks that are
implied by economic theory.
In our application, we set q = 4 and identify four factors: global activity, the real
price of oil, Norwegian oil specific activity, and Norwegian non-oil (Mainland) activity.
The number of factors and names are motivated by the model as discussed in Section 3
above.11 Of these four factors, the first two are observable and naturally load with one on
the corresponding element in the yt vector. The two latter factors must be inferred from
the data. We require that the Norwegian oil specific activity factor loads with one on
value added in the petroleum sector, and the Norwegian Mainland activity factor loads
with one on value added in Mainland Norway. Note that while this identifies the factors,
it does not mean that the factors and the observables are identical as we will use the full
information set to extract the factors.
Based on a minimal set of identification restrictions, we identify four structural shocks:
a global demand shock, an oil specific shock, a Norwegian oil activity shock (energy booms)
and a Norwegian non-oil (domestic) activity shock. The shocks are identified by imposing
a recursive ordering of the latent factors in the model, i.e. ft = [ftgact , ftoilp , ftoact , ftnoact ]0 ,
11
Moreover, as we show in Appendix C.1, four factors also explain a large fraction of the variance in the
dataset.
12
such that Q = A0 A00 . Specially, the mapping between the reduced form residuals ut and
structural disturbances et , ut = A0 et , is given by:
 gact  
  gdem 
ut
a11 0
0
0
et
 oilp  
  oils 
0   et 
 ut  a21 a22 0
(7)
 oact  = 


 ut  a31 a32 a33 0   eoact

t
a41 a42 a43 a44
unoact
enoact
t
t
where eit are the structural disturbances for i = [gdem, oils, oact, noact], with et e0t = I,
and [gdem, oils, oact, noact] denote global demand, oil specific, Norwegain oil activity and
non-oil activity, respectively.
For most energy importing countries, a higher price of oil causes production costs and
inflation to gradually increase, thereby eventually affecting overall activity. We therefore
follow the usual assumption from both theoretical and empirical models of the oil market, and restrict global activity to respond to oil specific disturbances with a lag. This
restriction is consistent with the sluggish behavior of global economic activity after each
of the major oil price increases in recent decades.
Furthermore, any unexpected news regarding global demand is assumed to affect the
real price of oil contemporaneously. As such, and consistent with recent work, we do
not treat the real price of oil as exogenous to the rest of the macro economy, see, e.g.,
Aastveit et al. (2012). In doing so, we confirm that both global demand and the oil
specific shock can drive up oil prices significantly. However, whereas the global demand
shock also stimulates global activity, the oil specific shock reduces global activity (with a
lag) and can therefore be interpreted as an adverse supply shock to the oil market.
In the short run, disturbances originating in the Norwegian economy are exogenous
to global activity and the real oil price. These are plausible assumptions, as Norway is a
small open economy that only accounts for less than three percent of global oil production.
However, both the oil and the non-oil domestic activity factors respond to unexpected
disturbances in global activity and the real price of oil on impact. In a small open economy
such as Norway, news regarding global activity will affect variables such as the exchange
rate, the interest rate, asset prices and consumer sentiments contemporaneously, and
thereby affect overall demand in the economy. Norway is also a net oil exporter. Thus,
any disturbances to the real price of oil will most likely rapidly affect both the demand
and supply side of the economy.
Lastly, in the short run, the oil activity factor is exogenous to the rest of the domestic
economy but can affect the other sectors contemporaneously (for instance via productivity
spillovers). However, and as discussed in Section 3, after a period we allow the energy
sector to respond to the dynamics in the other sectors of the economy.
4.2
Estimation
Let ỹT = [y1 , · · · , yT ]0 and f˜T = [f1 , · · · , fT ]0 , and define H = [λ0 , · · · , λs ], β = [φ1 , · · · , φh ],
Q, R, and pi = [ρ1,i , · · · , ρl,i ] for i = 1, · · · , N , as the model’s hyper-parameters.
Inference in our model can be performed using both classical and Bayesian techniques.
In the classical setting, two approaches are available, two-step estimation, and maximum
13
likelihood estimation (ML). In the former, f˜T , H and R are first typically estimated using
the method of principal components analysis (PCA), then the dynamic components of the
system, A and Q, are estimated conditional on f˜T , H and R. Thus, the state variables are
treated as observable variables. If estimation is performed using ML, the observation and
state equations are estimated jointly. However, employing ML still involves some type
of conditioning. That is, we first obtain ML estimates of the model’s unknown hyperparameters. Then, to estimate the state, we treat the ML estimates as if they were the
true values for the model’s nonrandom hyper-parameters. In a Bayesian setting, both the
model’s hyper-parameters and the state variables are treated as random variables.
We estimated the DFM using both the two-step procedure in the classical setting and
Bayesian estimation. The results reported in section 5 are not qualitatively affected by
the choice of estimation method. However, we prefer the Bayesian approach primarily due
to: 1) In contrast to the classical approach, inferences regarding the state are based on the
joint distribution of the state and the hyper-parameters, not a conditional distribution.
2) ML estimation would be computationally intractable given the number of states and
hyper-parameters. 3) Our data are based on logarithmic year-on-year differences. This
spurs autocorrelation in the idiosyncratic errors.
In a Bayesian setting, the model can readily be extended to accommodate these features of the error terms. In a classical two-step estimation framework, this is not the case.
Furthermore, in the two-step estimation procedure, it is not straightforward to include
lags of the dynamic factors in observation equation.
Thus, our preferred model is a Bayesian Dynamic Factor Model (BDFM). We set,
s = 2, h = 8, and l = 1. That is, we include 2 lags for the dynamic factors in the
observation equation (see equation 1), 8 lags in the transition equation (see equation 2),
and let the idiosyncratic errors follow AR(1) processes (see equation 4).12 In section C.1
we explain the choice of this particular specification and analyze its robustness.
4.2.1
The Gibbs sampling approach
Bayesian estimation of the state space model is based on Gibbs simulation, where the
following three steps are iterated until convergence is achieved:
Step 1: Conditional on the data (ỹT ) and all the parameters of the model, generate f˜T
Step 2: Conditional on f˜T , generate β and Q
Step 3: Conditional on f˜T , and data for the i-th variable (ỹT,i ), generate Hi , Ri and pi
for i = 1, · · · , N
In Appendix D we describe each step in greater detail and document the employed
prior specifications. We simulate the model using a total of 50000 iterations. A burn-in
period of 40000 draws is employed, and only every 5th iteration is stored and used for
inference.13
12
13
Note that we let s = 0 and l = 0 when estimating the DFM using the two-step estimation procedure.
Standard MCMC convergence tests confirm that the Gibbs sampler converges to the posterior distribution. Convergence statistics are available on request.
14
5
Results
Our results are presented in the following subsections. We first present the identified
factors before investigating how GDP, investment, employment and wages in the Mainland
economy and the real exchange rate respond to the various shocks. Then we examine the
sectoral reallocation following the energy booms and oil price shocks, before investigating
the implications for spending in the public sector in greater detail.
5.1
Factors and global shocks
The upper panel of Figure 3 displays, from the left, the global activity factor, the real
price of oil, the oil activity factor and the non-oil (Mainland) activity factor. The two
first factors are treated as observables in the estimation. Accordingly, they are measured
without uncertainty.
Global activity declined during the Asian crisis in the latter part of the 1990s, following
the dot com bubble that burst in 2000/2001, and during the recent financial crisis. The
latter trough is by far the most severe. Turning to the real oil price, Figure 3 suggests
that the most pronounced cycles in the real price of oil follow global activity cycles. There
is significant growth in the real oil price during the economic booms in 1999/2000 and
2006/2007 and a decrease in the real price of oil during the Asian crisis and the recent
financial crisis.
It is more interesting to investigate the cyclical patterns of the estimated latent factors,
i.e., oil activity and non-oil activity. Statistically, both factors are identified. As seen in
the figure, they are also economically meaningful. The latent oil activity factor shows
booms and busts that relate to the petroleum sector, such as the investment boom in
the North Sea in the middle of the 1990s, the decline in activity from 2000 (when oil
production peaked) and the decline in new investments in the period after the financial
crisis. The non-oil factor shows cyclical patterns that are well in line with the conventional
view of the Norwegian business cycle over the last two decades. The bust in 2002/2003, the
subsequent boom, and the recent bust during the financial crisis stand out. As expected,
the volatility of the oil activity factor is larger than that of the non-oil activity factor.
The estimation procedure we employ, see section 4, is inherently a smoothing algorithm. Thus, it is unsurprising that the oil and non-oil activity factors resemble the
cyclical patterns of oil investment (cyclical contribution) and the GDP of Mainland Norway, respectively, both displayed in Figure 1d. Importantly, however, the factors and the
observables are not identical. As stressed in section 2, the oil sector’s contribution to
the domestic economy comes through many more channels than investments alone. The
information set used to extract the two latent factors reflects this, as do the estimated
factors.
As discussed in section 3.1 above, we do not wish to treat the oil price as exogenous
and allow for reverse causality from global activity to the oil price. This implies that both
supply and demand shocks can affect oil prices. Figure 3, lower panel, illustrates this. It
displays the effect of a global demand shock to global activity and the real oil price and
subsequently the effect of an oil specific shock to the same two variables. While the global
15
Figure 3. Factors and global impulse responses
Factors
Real oil price
Global activity
0.04
Oil activity
1
0.1
0.8
0.02
0.04
0.6
0.05
0.4
0
Non-oil activity
0.06
0.02
0.2
−0.02
0
0
0
−0.02
−0.2
−0.04
−0.05
−0.4
−0.04
−0.6
−0.06
−0.08
1996.01
2000.02
2004.03
2008.04
2012.04
−0.06
−0.1
−0.8
−1
1996.01
2000.02
2004.03
2008.04
2012.04
1996.01 2000.02 2004.03 2008.04 2012.04
−0.08
1996.01
2000.02
2004.03
2008.04
2012.04
Global impulse responses
Global demand shock
Global act. resp.
Oil price resp.
Oil specific shock
Global act. resp.
Oil price resp.
10
0.1
1
10
0
5
−0.1
0
0
0
−0.2
−0.3
−1
−10
−5
−0.4
5
10
15
20
25
30
35
5
10
15
20
25
30
35
5
10
15
20
25
30
35
5
10
15
20
25
30
35
Note: The first row in the figure displays the observed variables and the estimated latent factors. The
second row displays impulse responses. The responses are displayed in levels of the variables. The Global
demand shock is normalized to a 1 percent increase, while the oil specific shock is normalized to increase
the real price of oil with 10 percent. The black solid lines are median estimates. The gray shaded areas
are 68 percent probability bands.
demand shock increases both activity and the real oil price, the oil specific shock generates
a temporary inverse relationship between the oil price and global activity, equivalent to
a supply type disturbance. Again, this is consistent with recent studies that have found
that a large fraction of the variation in the real price of oil can be attributed to global
demand, see e.g. Lippi and Nobili (2012) and Aastveit et al. (2012) among many others.
5.2
A resource rich economy
Table 1 displays the variance decomposition to the four identified shocks: oil activity
(energy booms), oil specific, global demand and non-oil activity, for GDP, employment,
investment and wages in the oil sector, the non-oil sector (Mainland Norway) and the
public sector, as well as for the real exchange rate. Figure 4 then displays the impulse
responses to the four identified shocks for the mainland economy and the real exchange
rate.
As expected, the oil activity and oil specific shocks together explain 60-70 percent of
the variation in production, employment, wages and investment in the petroleum sector.
However, while the investment dynamics in the petroleum sector are strongly associated
with oil specific shocks (that drive up oil prices), oil activity shocks are most important
for value added and employment. Lastly, global demand shocks (that drive up oil prices)
also affect the oil sector, and in particular petroleum investment. More than 20 percent
of the variation in petroleum investment refers back to global demand and its effect via
16
Table 1. Variance decompositions
Shock
Variable
Sector
&
Horizon
Oil
activity
Oil
specific
Global
demand
Non-oil
activity
4, 8
4, 8
4, 8
4, 8
GDP
Oil
Mainland
Public
0.82, 0.69
0.25, 0.32
0.06, 0.05
0.13, 0.12
0.06, 0.04
0.48, 0.40
0.04, 0.13
0.49, 0.44
0.01, 0.05
0.02, 0.06
0.20, 0.20
0.45, 0.50
Employment
Oil
Mainland
Public
0.66, 0.54
0.08, 0.04
0.21, 0.15
0.24, 0.20
0.12, 0.16
0.18, 0.23
0.06, 0.12
0.20, 0.28
0.05, 0.08
0.04, 0.14
0.59, 0.52
0.56, 0.54
Wages
Oil
Mainland
Public
0.46, 0.41
0.19, 0,08
0.66, 0.37
0.36, 0.29
0.05, 0.08
0.08, 0.15
0.15, 0.17
0.26, 0.38
0.05, 0.15
0.03, 0.13
0.49, 0.47
0.21, 0.32
Other
Investment Oil
Investment Mainland
Real Exchange Rate
0.01, 0.03
0.17, 0.28
0.11, 0.22
0.74, 0.61
0.28, 0.16
0.67, 0.58
0.21, 0.20
0.49, 0.49
0.23, 0.20
0.04, 0.15
0.06, 0.06
0.00, 0.00
Note: Each row-column intersection reports median variance decompositions for horizons 4 (left) and 8
(right)
higher oil prices.
What are the implications for the rest of the economy? Clearly, the oil boom stimulates
the mainland economy. In particular, Figure 4 shows that a boom in the energy sector
that increases oil activity by one percent increases GDP and investment in the mainland
sector by 0.4 and 0.7 percent, respectively, after 1-2 years. The effect is substantial;
approximately 30 percent of the variation in each of these variables is explained by energy
booms (see Table 1).
The spillovers from the energy sector to the labor market are more gradual. Employment and wages eventually increase after a year, peaking after 2-3 years. Ultimately,
energy booms are more important for wage dynamics than for employment, explaining
more than 20 percent of the changes in wages versus less than 10 percent of the employment variation in the mainland economy. The evidence is consistent with the view that
productivity increases in the energy sector worked to raise labor income in all sectors via
the centralized system of pay determination.
Lastly, the response in the real exchange rate is small and mostly insignificant, if
anything, showing evidence of real depreciation. This helps to explain why energy booms
can have such stimulative effects on the mainland economy.
There are two structural shocks that increase oil prices, an oil specific shock and a
global demand shock. Figure 4 shows that an oil specific shock is strongly associated
with real exchange rate appreciation. In fact, 60-70 percent of the variation in the real
exchange rate is explained by oil specific shocks, see Table 1. However, after 2-3 years,
the currency appreciation effect no longer operates.
17
Figure 4. Domestic impulse responses
GDP Mainland Norway
Oil act. shock
Oil spec. shock
Global dem. shock
Non-oil act. shock
0.6
0.6
1.5
1
0.4
1
0.3
0.5
0.2
0
0
0
0
−0.3
−0.2
−0.5
−1
−1
5
10
15
20
25
30
35
−0.6
5
10
15
20
25
30
5
35
10
15
20
25
30
35
5
10
15
20
25
30
35
Investment Mainland Norway
Oil act. shock
Oil spec. shock
Global dem. shock
Non-oil act. shock
2
1
1
2
1
0.5
0.5
0
0
0
−1
0
−2
−0.5
−2
−0.5
5
10
15
20
25
30
35
5
10
15
20
25
30
35
5
10
15
20
25
30
35
5
10
15
20
25
30
35
Employment Mainland Norway
Oil spec. shock
Oil act. shock
Global dem. shock
Non-oil act shock
1
1.5
0.4
0.2
0.5
1
0.2
0
0
0.5
−0.5
0
0
−0.2
−1
−0.5
−0.2
5
10
15
20
25
30
35
5
10
15
20
25
30
35
5
10
15
20
25
30
35
5
10
15
20
25
30
35
Wages Mainland Norway
Oil spec. shock
Oil act. shock
Global dem. shock
Non-oil act. shock
2
0.6
2
0.4
1
0.5
0.2
1
0
0
0
0
−0.2
−1
−1
−0.4
5
10
15
20
25
30
5
35
10
15
20
25
30
−2
35
5
10
15
20
25
30
5
35
10
15
20
25
30
35
Real exchange rate
Oil act. shock
Oil spec. shock
Global dem. shock
Non-oil act. shock
2
0.2
1
1
0
1
0.5
0.5
−0.2
−0.4
0
0
0
−0.6
−1
−0.5
−0.5
−0.8
5
10
15
20
25
30
35
5
10
15
20
25
30
35
5
10
15
20
25
30
35
5
10
15
20
25
30
35
Note: The responses are displayed in levels of the variables. All shocks are normalized to a 1 percent
increase, except for the oil specific shock, which is normalized to increase the real price of oil with 10
percent. The gray shaded area represent 68 percent probability bands, while the black solid lines are
median estimates.
18
Table 2. Productivity
Shock
Oil activity
Oil specific
Horizon
4
8
16
0.36
0.01
0.25
-0.01
0.22
-0.00
Note: The numbers show the difference between the response in value added and employment for Mainland,
interpreted as labour productivity.
The oil specific shock also has spillovers to the rest of the economy, although to a lesser
extent than the oil activity shock. In particular, following an oil specific shock that
increases oil prices by 10 percent, GDP and investment in Mainland Norway increase
temporarily by 0.25 and 1 percent, respectively, most likely as petroleum investment also
increases, see Table 1. Furthermore, employment and wages gradually increase, suggesting
that there are spending effects owing to the windfall gains associated with increased oil
prices.
The second shock that can potentially increase oil prices, a global demand shock, also
causes the Norwegian currency to appreciate. However, the response in the exchange rate
is less pronounced than for the oil specific shock, explaining approximately 20 percent of
the real exchange rate variation. As a consequence, the effect on GDP and investment,
as well as the spillovers to employment and wages, are more substantial. Between 40 and
50 percent of the variation in mainland GDP and investment activities can be explained
by global demand.14 The finding that foreign factors are important for the Norwegian
business cycles is consistent with Aastveit et al. (2011) and Furlanetto et al. (2013).
Lastly, a non-oil (domestic) activity shock increases GDP, employment and wages
in the mainland economy. The effect on investment is also positive, but the variation
explained by the domestic shock is modest (less than 10 percent). The effect on the real
exchange rate is negligible.
It is too early to make any conclusions regarding any evidence (or lack thereof) of
Dutch disease. To do so, we need to examine sectoral reallocation, which we do below.
However, it is obvious that the Norwegian economy has benefitted from having a highly
profitable oil and gas sector: Both windfall gains due to energy booms and higher oil prices
had positive spillover effects on the mainland economy. What are the mechanisms behind
these spillovers? While we have seen that labor input clearly increased following this
shock, Table 2, which measures productivity gains after 4, 8 and 16 quarters, suggests that
productivity spillovers are also of first order importance for energy booms. As productivity
measures the efficiency of production, this also explains why investment in the mainland
economy increased substantially following this shock. This is interesting, as it highlights
the empirical relevance of alternative theoretical Dutch disease models, see, e.g., Torvik
14
An one-percent increase in global demand, increases real oil prices by approximately 10-12 percent, see
Figure 3. Compared to a similar sized oil price increase due to an oil specific shock, the effects on GDP
and investment in Mainland Norway are more than twice as large; GDP increases by 0.7-1 percent after
a year, while investment increases by 2 percent.
19
Table 3. Residual regressions
Shock
Variable
R2
Lag
1
2
3
4
Oil activity
CPI
PPI
OSEBX
* Energy
ToT
-0.00 (0.55)
-0.03 (0.01)
0.08 (0.27)
0.13 (0.10)
-0.01 (0.56)
-0.00 (0.92)
0.01 (0.69)
0.11 (0.03)
0.14 (0.00)
0.01 (0.22)
-0.00 (0.52)
0.00 (0.98)
0.10 (0.19)
0.06 (0.39)
0.00 (0.95)
0.00 (0.25)
-0.01 (0.20)
-0.02 (0.67)
-0.06 (0.49)
0.00 (0.76)
0.02
0.06
0.16
0.17
0.02
Oil specific
CPI
PPI
OSEBX
* Energy
ToT
0.00
0.04
0.09
0.12
0.01
0.00
0.02
0.03
0.07
0.01
0.00
0.03
0.03
0.07
0.01
0.00 (0.24)
0.00 (0.81)
-0.05 (0.31)
0.00 (0.95)
-0.01 (0.21)
0.15
0.34
0.13
0.19
0.16
(0.60)
(0.00)
(0.02)
(0.01)
(0.02)
(0.13)
(0.00)
(0.50)
(0.23)
(0.17)
(0.01)
(0.00)
(0.48)
(0.14)
(0.22)
Note: For each variable the rows show coefficient estimates and Newey-West estimated p-values (in
parenthesis) from simple OLS regressions:
yt,i = αi,j +
P
X
βp,i,j et−p,i,j + ut,i,j
p=1
where i denotes variable i = 1, · · · , 4, j denotes structural shocks j = [Oil activity and Oil specific],
and p are the number of lags with P = 4. All dependent variables, yt,i , are transformed to four quarter
logarithmic differences. et−p,i,j are the median estimates of the structural shocks. The sample is 1997Q1−
2012Q4.
(2001), which emphasize learning by doing mechanisms and productivity spillovers.
Conversely, the oil specific shocks (that increase oil prices) have virtually no effect
on productivity, see Table 2. As such, our results show that is important to distinguish
between windfall gains due to volume and price changes when analyzing the Dutch disease
hypothesis. To the best of our knowledge, this is the first paper to explicitly separate and
quantify these two channels, while also allowing for explicit disturbances to world activity
and the non-oil sector.
Table 3 adds further evidence to the structural interpretation. In the table, we separately regressed the lags of the median structural shocks on consumer price inflation
(CPI), producer price inflation (PPI), total stock returns (OSEBX), stock returns for the
energy firms (Energy) and the terms of trade (ToT). Although simple, these regressions
not only confirm that the structural identification of our benchmark model is sound but
also shed light on the additional channels through which the energy sector affects the
economy.
First, as asset prices are the present discounted values of the future net earnings of the
firms in the economy, unexpected energy booms that enhance the production possibilities
for the whole economy should be positively related to stock returns. This is confirmed in
our regressions, where the oil activity shock explains a considerable share of the variation
in stock returns (both OSEBX and Energy). We find no evidence that the shock increases
costs, as energy booms do not explain a substantial amount of the variation in CPI and
20
PPI. Furthermore, the effect on terms of trade is insignificant, confirming that the windfall
gains associated with energy booms are not related to energy prices. Instead, energy
booms change the distribution of wealth due to productivity spillovers, the subsequent
movement of resources, higher income and increased spending in the overall economy.
Moreover, we find that the oil specific shock leads to a general rise in production costs
(PPI). This erodes the real effect of spending and may explain why this shock has less
stimulating effects on the economy. However, we confirm that the terms of trade are
positively affected by oil price increases in an oil exporting economy, which explains the
pronounced effect on the real exchange rate we observed above. Furthermore, oil specific
shocks also explain a substantial share of the variation in energy specific stock returns.15
The results presented thus far reflect average responses over the sample analyzed. In
Figure 5, we show that the structural shocks are also well identified in terms of timing. In
particular, the figure displays the model’s historical decomposition of the domestic factor
representing the non-oil economy. As seen in the figure, oil activity shocks stimulated the
Norwegian economy, particularly from the middle of the 1990s and until 2000 (after which
there was a temporary cyclical decline in oil activity, see also Figure 1d), and again during
the economic upswing beginning around 2004. However, while the period of high economic
growth in the middle and late 1990s can in large part be explained by increased oil activity,
the high growth period predating the financial crisis was primarily driven by increased
global demand and oil specific shocks, which both drove up oil prices. The windfall gain
from higher oil prices stimulated investment in the petroleum sector and thereby also
the mainland economy through spillover effects. However, by the end of 2008, Norway
was affected by the financial crisis. The subsequent downturn was primarily caused by
negative global demand as well as by oil specific shocks (that lowered oil prices) and oil
activity shocks. The return to trend growth was gradual, with positive contributions
from oil specific shocks. From 2011, global demand again contributed positively to the
mainland economy (again via higher oil prices).
For the reader with detailed knowledge of the Norwegian economy, Figure 5 presents
a reasonable story of a country that has benefited from increased activities in the North
Sea, albeit with cyclical up and downturns. However, the negative or only mildly positive contribution from the oil activity shocks since 2006/2007 provides some cause for
concern. To the extent that an oil boom is associated with productivity dynamics (that
positively affect value added in the overall economy), the muted role of these shocks suggests that that productivity spillovers have declined recently. This is consistent with the
view portrayed in Olsen (2013) of a slow down in productivity since 2005. Furthermore,
labor input per hour worked has also declined in recent years relative to Norway’s trading
partners. Thus, while the enhanced linkages from both the oil sector and energy prices
have been positive for growth and employment in the Norwegian economy for nearly two
decades, the declining productivity spillovers coupled with increased costs could be a
15
Without reading too much into these simple regressions, we also observe that the oil specific shocks
explain more of the energy specific returns than overall returns in the economy (measured by OSEBX),
which is consistent with the view that the increased costs eroded the value added from the oil specific
shock. I.e., while firms in the Energy sector benefit from increased oil prices, the spillover to value added
in the overall economy was small.
21
Figure 5. Historical shock decomposition: Non-oil activity Norway
Global demand
Oil specific
Oil act.
Non−oil act.
0.02
0
−0.02
−0.04
−0.06
1996.01
2000.02
2004.03
2008.04
2012.04
Note: The figures report the accumulated contribution of each structural shock to the growth in the non-oil
activity factor.
major concern in the long run.
5.3
Sectoral performance - Two speed boom?
Figure 6 displays the responses in value added and employment to energy booms (left
column) and oil specific shocks (right column). The figure displays the quarterly average
of each sector’s response (in levels) to the different shocks. The oil activity shock is
normalized to increase oil activity by 1 percent, while the oil specific shock increases oil
prices by 10 percent (which is customary in the literature). Note that the white bars
indicate that the shock explains less than 10 percent of the variation in a sector.
The figure emphasizes that energy booms stimulate value added in all industries in
the private sector, but to a varying degree. The construction and business sectors are
among the most positively affected. Between 30 and 40 percent of the variance in these
sectors is explained by energy booms, see Table 4 in Appendix B. These are industries
with moderate direct input into the oil sector, but the indirect effects are large. Value
added in manufacturing is also positively affected, but less so than in the non-tradable
sectors. Yet, there is no evidence of Dutch disease wherein the sector eventually contracts.
Turning to the labor market, our model confirms the stylized facts presented above in
Figure 1c. Norway has become a two speed economy, with employment in non-tradable
sectors such as construction, the business service sector and real estate growing at a much
faster pace than tradables such as manufacturing. However, and as above, there is no
evidence of Dutch disease; manufacturing does not contract. Interestingly, the effect on
the public sector (value added and employment) is negligible, suggesting only a minor
government spending effect following this shock.
Are these numbers reasonable? Compared to Eika et al. (2010), who calculate the
direct and indirect effects based on input-output tables, our numbers are more substantial.
Yet, Eika et al. (2010) also found the service sector (e.g., as business industries) to be
the most affected, once accounting for indirect effects such as inputs between the sectors.
Where we diverge is in the size of the spillovers and the number of sectors involved.
22
Figure 6. Relative responses
Value added: Oil activity shock
Value added: Oil specific shock
Employment: Oil activity shock
Employment: Oil specific shock
Note: Each plot displays the quarterly average of each sector i’s response (in levels) to the different shocks.
The averages are computed over horizons 1 to 12. The oil activity shock is normalizes to increase oil
activity by 1 percent, while the oil specific shock is normalized to increase the real price of oil with 10
percent. White bars indicate that the shock explains less than 10 percent of the variation in the sector.
However, this should come as no surprise, as we also allow for induced spending effects
via income and wage growth, see Table 1. Moreover, in our framework the input-output
table becomes endogenous, as we allow for shared productivity dynamics across sectors.
As seen in Table 1, and indicated by the white bars in Figure 6, the oil specific shock
generally explains a substantially smaller share of the variance in the sectoral variables
than the oil activity shock. The responses to the oil specific shock also present a more
diverse picture. Now sectors such as scientific services and manufacturing are among the
most positively affected. This is interesting, as these sectors are also technology intensive
and enjoy spillovers from the significant boost in petroleum investment that follows the oil
specific shock. As offshore oil often demands complicated technical solutions, the oil specific shock generated positive knowledge externalities that benefited employment in these
sectors in particular. Thus, the theory of Dutch disease is turned on its head following
this shock. However, compared to the responses reported for the oil activity shock, the
public sector is now also positively affected, suggesting the presence of a spending effect.
We examine this in greater detail in the next section.
Lastly, the global demand shock is important for all industries in the private sector, but
23
most so for manufacturing (relative plots are available on request).16 Thus, the stylized
fact that manufacturing is lagging behind the other sectors, in particular in the financial
crisis (see Figure 1c), also refers to manufacturing’s substantial exposure to foreign shocks
(which were all negative in the financial crisis).
In summary, we find no evidence of Dutch disease as experienced in the Netherlands
in the 1970s. Instead, we find positive spillovers between the energy sector and both the
tradable and non-tradable sectors. As discussed, an important channel for these spillovers
could be productivity and learning by doing. As such, our results highlight the empirical
relevance of alternative theoretical Dutch disease models, such as that proposed by, e.g.,
Torvik (2001). Moreover, our model successfully replicates the stylized facts portrayed in
Figure 1 indicating a two speed economy. Importantly, however, the observed two speed
pattern is not a function of resource wealth in isolation; global factors need to be taken
into account.
5.4
Public sector
One aspect of the results presented above that we have not discussed in detail thus far
is the role of the public sector. Norway has a large public sector, and much of the
petroleum income is directly managed through the Norwegian Petroleum Fund, which
was specially designed with the express purpose of shielding the domestic economy from
potential spending effects caused by the resource endowment. Through a fiscal rule, which
permits the government to spend approximately 4 percent of the fund (expected return)
every year, the income from the oil and gas sector should only gradually be phased into
the economy, and thus ensure fiscal discipline.
Very few studies have analyzed the effects of oil price changes on government spending
in Norway. Those that do find very small effects, see, e.g., Pieschacon (2012). However,
Pieschacon (2012) does not control for the different sources that may affect the oil price.
As we have shown here, oil price increases can be due to either global demand or oil specific
shocks, and the mechanisms by which they affect the economy will not be identical.
Although we do not explicitly examine fiscal policy in this study, the results presented
above reveal two interesting points regarding government spending in a resource rich
economy. First, energy booms do not explain a large share of the variance in value added
or employment in the public sector. As such, governmental arrangements to ensure fiscal
discipline seem to work.17 However, the results presented in Figure 6 suggested that the
public sector is positively affected by the oil specific shock. Furthermore, 40 percent of
the variation in government spending is explained by oil specific shocks (see Table 1).
This suggests evidence of a spending effect from increased oil prices via the public sector,
even though the fiscal rule is in place. To explore this further, we augment the dataset
with the value added in the central and local governments, and re-estimate the model.18
16
As seen in tables 1 and 4, the variance explained by the global demand shock is substantial for all sectors
except the public one. However, the manufacturing sector is by far the most affected; 60 percent of the
variation in value added in the manufacturing sector can be explained by foreign shocks.
17
However, if the oil activity shocks are pure productivity spillovers, the public sector does not seem to
benefit from these in the same manner as the other sectors of the economy.
18
The baseline results are not quantitatively affected.
24
Figure 7. Impulse responses: Oil specific shocks
Government total
Central government
0.03
0.03
0.02
0.02
Local government
0.05
0.01
0.01
0
0
0
−0.01
−0.05
−0.01
5
10
15
20
25
30
35
−0.02
5
10
15
20
25
30
35
5
10
15
20
25
30
35
Note: The responses are displayed in levels of the variables. The oil specific shock is normalized to
increase the real price of oil with 10 percent. The gray shaded area represent 68 percent probability bands,
while the black solid lines are median estimates.
Norway has had an active population maintenance policy for rural districts. We therefore
expect the increased income from the North Sea to have benefited local governments in
particular. The impulse responses of the newly added government variables are displayed
in Figure 7.
The results emphasize that there is a positive link between increased oil prices and
government spending, in particular at the local government. While 30-40 percent of the
variation in value added at the state level is determined by oil specific shocks, the corresponding number for local government is close to 60 percent. The results are consistent
with Norway having an active government policy of investing in rural development. These
could be regions that may not directly benefit from oil-related developments.
On a final note, in our model, government spending will respond to the various shocks
affecting the economy. As oil specific shocks are generally beneficial for an oil exporter such
as Norway, but less so for other oil importing countries (see Figure 3), increased spending
by the government could also be a way to shelter the economy from a decline in foreign
demand due to higher oil prices. However, our analysis shows that the consequences of
increased spending will be manifested in an appreciated exchange rate and eventually
increased costs, as can be seen in Figure 6 and Table 3, respectively. Both mechanisms
deteriorate competitiveness, which could be a concern for the Norwegian economy in the
long run.
6
Additional results and robustness
As mentioned in the main text, our results are robust to estimating the model using
classical two-step estimation techniques. Furthermore, as descried in Section D.0.4, our
results seem robust to different prior specifications. We have also conducted a series of
other robustness checks. These are described fully in Appendix C. Below, we provide a
brief summary.
First, global activity is not observed. We have approximated global activity by taking
the simple mean across eight countries thought to be important to the global business
cycle and Norway in particular. Qualitatively, the results reported in Section 5 are not
25
affected by excluding countries from this set or changing its composition. Details are
provided in Appendix C.2.
Second, running the analysis on a different sample does not change the main conclusions reported in Table 1. That is, on average across sectors, a booming oil sector explains
approximately 20-30 percent of the variation in the disaggregated series, irrespective of
whether we estimate the model on a sample period from 1986:Q1 to 2012:Q4 or 1996:Q4
to 2012:Q4. However, as described in Appendix C.3, the subsample analysis should be
interpreted with caution due to differences in data availability.
Third, the model specification is uncertain. The number of factors and lags employed
in the model should be tested. We do this primarily by running a quasi-real-time forecasting experiment. The results reported in Appendix C.1 show that our benchmark model,
outlined in Section 4, performs superior to simple univariate autoregressive processes.
The Benchmark specification is also among the best performing specifications and is the
best model specification over shorter forecasting horizons.
7
Conclusion
This study examines the empirical validity of the classical Dutch disease theory in a small
and open oil and gas producing economy. Using Norway as a case study, we provide a
novel contribution on the subject by explicitly identifying and quantifying windfall gains
from a booming energy sector or higher oil prices and the associated sectoral performance
in the rest of the economy.
We estimate a Bayesian Dynamic Factor Model that includes separate activity factors
for oil and non-oil sectors in addition to global activity and the real price of oil. The
model is particularly useful in a data rich environment such as ours, where common latent
factors and shocks are assumed to drive the co-movements between economic variables in
the economy.
We have two main results: First, booms in the energy sector have substantial productivity spillovers on the non-oil sectors, effects that have not been captured in previous
analysis. In particular, we find that the energy sector stimulates investment, production,
employment and wages in nearly all non-oil industries. Construction, business services
and real estate are the most stimulated sectors. Second, windfall gains due to changes in
the real oil price also stimulate the economy, but primarily if the oil price increase relates
to a boom in global demand. Oil price increases due to, say, supply disruptions, while
stimulating activity in the technologically intense service sectors and boosting government
spending, have small spillover effects to the rest of the economy, in part because of a substantial real exchange rate appreciation and reduced cost competitiveness. Yet, there is
no evidence of Dutch disease as experienced in the Netherlands in the 1970s, where natural gas discoveries had adverse effects on the Dutch manufacturing sector. Instead, there
is evidence of a two speed economy, with the manufacturing sector lagging behind the
booming service sectors. Importantly, however, the observed two speed pattern is not a
function of resource wealth in isolation; global factors need to be taken into account.
Our results suggest that traditional Dutch disease models with a fixed capital stock
26
and exogenous labor supply do not provide a convincing explanation for how petroleum
wealth affects a resource rich economy when there are productivity spillovers between the
various sectors.
27
References
Aastveit, K. A., H. C. Bjørnland, and L. A. Thorsrud (2011). The world is not enough!
Small open economies and regional dependence. Working Paper 2011/16, Norges Bank.
Aastveit, K. A., H. C. Bjørnland, and L. A. Thorsrud (2012). What drives oil prices?
Emerging versus developed economies. Working Paper 2012/11, Norges Bank.
Bai, J. and S. Ng (2002). Determining the number of factors in approximate factor models.
Econometrica 70 (1), 191–221.
Bai, J. and S. Ng (2010). Principal components estimation and identification of the
factors. Unpublished manuscript.
Bai, J. and P. Wang (2012). Identification and estimation of dynamic factor models.
MPRA Paper 38434, University Library of Munich, Germany.
Bjørnland, H. C. (1998). The economic effects of North Sea oil on the manufacturing
sector. Scottish Journal of Political Economy 45 (5), 553–85.
Boivin, J. and M. Giannoni (2006, December). DSGE Models in a Data-Rich Environment. Working Paper 12772, National Bureau of Economic Research.
Bruno, M. and J. Sachs (1982). Energy and resource allocation: a dynamic model of the
”Dutch Disease”. The Review of Economic Studies 49 (5), 845–859.
Carter, C. K. and R. Kohn (1994).
Biometrika 81 (3), pp. 541–553.
On gibbs sampling for state space models.
Charnavoki, V. and J. Dolado (2012). The effects of global shocks on small commodityexporting economies: New evidence from Canada. Discussion Paper DP8825, CEPR.
Corden, W. M. (1984). Booming sector and dutch disease economics: survey and consolidation. Oxford Economic Papers 36 (3), 359–380.
Corden, W. M. (2012). Dutch disease in australia: Policy options for a three-speed
economy. Australian Economic Review 45 (3), 290–304.
Corden, W. M. and J. P. Neary (1982). Booming sector and de-industrialisation in a small
open economy. The Economic Journal 92 (368), 825–848.
Eastwood, R. and A. J. Venables (1982). The macroeconomic implications of a resource
discovery in an open economy. The Economic Journal 92 (366), 285–299.
Eika, T., J. Prestmo, and E. Tveter (2010). Etterspørselen fra petroleumsvirksomheten.
Betydningen for produksjon og sysselsetting i Norge. Økonomiske analyser 3/2010,
Statisics Norway.
Furlanetto, F., S. Sarferaz, and F. Ravazzolo (2013). Business cycles in commodityexporting countries. Mimeo, Norges Bank.
28
Geweke, J. (1977). The dynamic factor analysis of economic time series. In latent Variables
in Socio-economic Models. D.J. Aigner and A.S. Goldberger (Eds.).
Hamilton, J. D. and A. M. Herrera (2004). Oil Shocks and Aggregate Macroeconomic
Behavior: The Role of Monetary Policy: Comment. Journal of Money, Credit and
Banking 36 (2), 265–86.
Hutchison, M. M. (1994). Manufacturing sector resiliency to energy booms: empirical
evidence from Norway, the Netherlands, and the United Kingdom. Oxford Economic
Papers 46, 311–329.
Ismail, K. (2010). The structural manifestation of the ”Dutch Disease”: The case of oil
exporting countries. IMF Working Papers, 1–36.
Kim, C.-J. and C. R. Nelson (1999). State-Space Models with Regime Switching: Classical
and Gibbs-Sampling Approaches with Applications (1 ed.), Volume 1. The MIT Press.
Kose, M. A., C. Otrok, and C. H. Whiteman (2003). International business cycles: World,
region, and country-specific factors. American Economic Review 93 (4), 1216–1239.
Kose, M. A., E. S. Prasad, and M. E. Terrones (2003). How does globalization affect the
synchronization of business cycles? American Economic Review 93 (2), 57–62.
Lama, R. and J. P. Medina (2012). Is Exchange Rate Stabilization an Appropriate Cure
for the Dutch Disease? International Journal of Central Banking 8 (1), 5–46.
Larsen, E. R. (2006). Escaping the Resource Curse and the Dutch Disease? American
Journal of Economics and Sociology 65 (3), 605–640.
Lippi, F. and A. Nobili (2012). Oil and the macroeconomy: A quantitative structural
analysis. Journal of the European Economic Association 10 (5), 1059–1083.
Mehlum, H., K. Moene, and R. Torvik (2006). Institutions and the Resource Curse. The
Economic Journal 116 (508), 1–20.
Mumtaz, H., S. Simonelli, and P. Surico (2011). International Comovements, Business
Cycle and Inflation: A Historical Perspective. Review of Economic Dynamics 14 (1),
176–198.
Neary, J. P. and S. van Wijnbergen (1984). Can an oil discovery lead to a recession? a
comment on eastwood and venables. The Economic Journal 94 (374), 390–395.
Olsen, Ø. (2013). Address by Governor Øystein Olsen to the Supervisory Council of Norges
Bank and invited guests on Thursday 14 February 2013. Economic Perspectives, Norges
Bank.
Pieschacon, A. (2012). The value of fiscal discipline for oil-exporting countries. Journal
of Monetary Economics 59 (3), 250–268.
29
Sachs, J. D. and A. M. Warner (1995). Natural resource abundance and economic growth.
Working papers 5398, National Bureau of Economic Research.
Stock, J. H. and M. W. Watson (2002). Macroeconomic forecasting using diffusion indexes.
Journal of Business & Economic Statistics 20 (2), 147–162.
Stock, J. H. and M. W. Watson (2005a). Implications of dynamic factor models for var
analysis. NBER Working Papers 11467, National Bureau of Economic Research, Inc.
Stock, J. H. and M. W. Watson (2005b). Understanding changes in international business
cycle dynamics. Journal of the European Economic Association 3 (5), 968–1006.
Thorsrud, L. A. (2013). Global and regional business cycles. Shocks and propagations.
Working Papers 0012, Centre for Applied Macro- and Petroleum economics (CAMP),
BI Norwegian Business School.
Torvik, R. (2001). Learning by doing and the Dutch disease. European Economic Review 45 (2), 285–306.
Van Wijnbergen, S. (1984). The ”Dutch Disease”: a disease after all? The Economic
Journal 94 (373), 41–55.
30
Appendices
Appendix A
Sector Abbreviation
Data and Sources
Moments
Variable in National Accounts
Mean
Std.
GDP
Oil extraction
Oil service
Manufacturing
Construction
Retail
Transp ocean
Transp mainland
Hotel and food
Financial
Real estate
Scientific
Business
Mainland
Public
-0.63
6.89
1.56
3.04
4.39
-5.32
1.03
1.76
3.90
9.20
4.12
6.79
2.79
1.64
6.28
25.65
3.53
5.02
3.46
16.22
5.26
4.84
7.88
10.81
4.75
6.91
1.95
1.46
Oil and natural gas extraction
Service activities incidental to oil and gas
Manufacturing
Construction
Wholesale and retail trade, repair of motor vehicles
Ocean transport
Transport activities excl. ocean transport
Accommodation and food service activities
Financial and insurance activities
Real estate activities
Professional, scientific and and technical activities
Administrative and support service activities
Mainland Norway
General government
Empl
Oil extraction
Oil service
Manufacturing
Construction
Retail
Transp ocean
Transp mainland
Hotel and food
Financial
Real estate
Scientific
Business
Mainland
Public
2.43
11.14
-0.59
3.95
1.28
0.91
0.70
1.10
-0.13
5.98
3.61
5.66
1.33
1.33
5.32
14.30
3.35
4.05
1.99
2.79
2.22
2.70
2.92
6.61
3.75
6.32
1.54
0.92
See above
Wages oil
Wages public
Wages mainland
Investment oil
Investment
mainland
Exchange rate
9.90
6.04
6.06
4.52
4.06
6.87
1.77
2.38
22.62
8.60
Wages petroleum
Wages public
Wages mainland
Investment petroleum
Investment mainland
0.57
4.79
BIS effective exchange rate index, broad basket
World activity
Oil Price
2.78
9.01
1.90
33.11
See text, section 3
Crude Oil-Brent, deflated using US CPI
Other
Int.
Note: The table lists all the variables used in the Benchmark model. All activity, investment, wages and
employment series for Norway are collected from the Quarterly National Accounts database of Statistics
Norway. The international series were downloaded from Datastream. The real exchange rate is collected
form BIS. All series are seasonally adjusted by their source. Std. denotes standard deviation. Int. denotes
international. The moments are computed based on the transformed variables, i.e. log(xi,t )−log(xi,t−4 )).
31
Appendix B
Figures and tables
Table 4. Variance decompositions
Shock
Variable
Sector
&
Horizon
Oil
activity
Oil
specific
Global
demand
Non-oil
activity
4, 8
4, 8
4, 8
4, 8
GDP
Manufacturing
Construction
Retail
Transport ocean
Transport mainland
Hotel and food
Financial
Real estate
Scientific
Business
0.04,
0.26,
0.74,
0.27,
0.21,
0.10,
0.19,
0.08,
0.07,
0.32,
0.08
0.35
0.81
0.37
0.32
0.19
0.31
0.14
0.03
0.42
0.27,
0.03,
0.01,
0.62,
0.18,
0.13,
0.58,
0.77,
0.43,
0.19,
0.18
0.01
0.02
0.48
0.10
0.06
0.43
0.70
0.34
0.09
0.65,
0.51,
0.21,
0.10,
0.55,
0.57,
0.22,
0.15,
0.41,
0.37,
0.69
0.46
0.15
0.11
0.51
0.54
0.24
0.15
0.50
0.36
0.04,
0.20,
0.03,
0.01,
0.06,
0.20,
0.01,
0.00,
0.08,
0.12,
0.04
0.18
0.02
0.04
0.06
0.20
0.02
0.00
0.12
0.13
Employment
Manufacturing
Construction
Retail
Transport ocean
Transport mainland
Hotel and food
Financial
Real estate
Scientific
Business
0.07,
0.28,
0.32,
0.10,
0.31,
0.29,
0.52,
0.10,
0.28,
0.04,
0.04
0.39
0.35
0.11
0.14
0.11
0.63
0.12
0.09
0.08
0.14,
0.14,
0.04,
0.34,
0.11,
0.14,
0.05,
0.22,
0.03,
0.16,
0.17
0.06
0.02
0.18
0.11
0.17
0.03
0.08
0.11
0.10
0.23,
0.40,
0.32,
0.22,
0.03,
0.30,
0.42,
0.52,
0.20,
0.61,
0.31
0.37
0.35
0.13
0.16
0.39
0.30
0.61
0.35
0.62
0.56,
0.18,
0.32,
0.34,
0.55,
0.28,
0.01,
0.16,
0.49,
0.19,
0.48
0.17
0.28
0.59
0.59
0.33
0.04
0.18
0.45
0.20
Note: Each row-column intersection reports median variance decompositions for horizons 4 (left) and 8
(right)
32
Appendix C
C.1
Robustness
Model specification
The correct model specification is uncertain. Different test statistics, see Bai and Ng
(2002), suggest between 3 and 8 static factors. Using 4 factors explains approximately 60
percent of the variation in the dataset. Including an additional 4 static factors increases
the variance explained by a modest 17 percent. Although informative, the tests for the
number of static factors are far from conclusive.
To fully test our preferred model relative to alternative specifications, we run a quasireal-time forecasting experiment. The experiment is conducted as follows: For the sample
period from 1996.01 to 2012.04, we estimate the BDFM with different lag specifications.
in particular, we allow for up to 2 lags of the vector of factors in the observation equation
(s = 0, . . . , 2). For each lag specification we also estimate the model with and without
autocorrelated idiosyncratic errors (l = 0, 1). Ultimately, this yields 6 different specifications. Lastly, for each of these combinations we estimate the model with 4 and 8 lags in
the transition equation (h = 4, 8).
We compute the model’s out of sample forecasting performance over the period from
1996.02 to 2012.04. The performance is scored by root mean forecasting errors (RMSE)
and log scores (logScore).19 The forecasting experiment is quasi-real-time, as we do not
re-estimate the models for each new vintage of data we forecast, and we also do not use
real-time vintage data when estimating the models or in the evaluation of forecasting
performance. Thus, the distribution of the model parameters used to forecast is assumed
to be constant throughout the evaluation period. For our purpose, which is to make
comparison among nested structural models, this is an innocuous assumption. Furthermore, an advantage of a quasi-real-time forecasting experiment, as opposed to a real-time
forecasting experiment, is that we can evaluate the forecasting performance over a much
longer sample.20
Table 5 reports the results.21 Panels (a) and (b) reports the results for h = 4 and
h = 8, respectively. At the two step ahead horizon, and evaluated across all variables, our
preferred model specification, BDFM s(2)a(1) (denoted Benchmark in the table), performs
substantially better than any other model specification. In Panel (a) (Panel (b)), for 20
(19) and 19 (21) out of 39 variables, the Benchmark model performs best in terms of
respectively RMSE and average logScore, respectively. At the four step ahead horizon,
the ranking of the different model specifications changes, and the BDFM s(2)a(0) model
receives a better score than the other models in approximately 40 to 50 percent of the
cases.
19
The RMSE is a quadratic loss function that is often used to evaluate point forecasts. If the focus is on
the whole forecast distribution, the RMSE is not appropriate and log scoring is a better metric. The
logScore is the logarithm of the probability density function evaluated at the outturn of the forecast. As
such it provides an intuitive measure of density fit.
20
I.e., in a real-time experiment, we would have to re-estimate the models for each new vintage and use a
substantial part of the sample to estimate the initial parameter distributions.
21
To save space, we only report the results for forecasting horizons 2 and 4. The conclusions do not change
for horizons 1 and 3. These results are available on request.
33
Table 5. Forecast performance
Horizon
Model
Variable
Y
All
E
I
W
variables
Panel a:
2
BDFM
BDFM
BDFM
BDFM
BDFM
AR(1)
4
s(0)a(0)
s(1)a(0)
s(2)a(0)
s(0)a(1)
s(1)a(1)
0.90,
1.01,
1.05,
0.99,
0.98,
0.77,
0.97
1.00
1.02
0.99
1.00
0.90
1.00,
1.02,
1.05,
0.97,
0.99,
0.62,
1.03
1.01
1.01
1.02
1.01
0.84
0.96,
0.98,
1.00,
0.97,
0.98,
0.86,
0.96
0.99
1.01
0.98
0.99
0.87
1.00,
1.06,
1.08,
1.00,
0.99,
0.51,
1.00
1.00
1.01
1.01
1.00
0.73
0, 1
1, 3
11, 7
0, 5
5, 2
2, 2
Benchmark
0.01, 2.92
0.01, 3.58
0.07, 1.30
0.01, 3.16
20, 19
BDFM
BDFM
BDFM
BDFM
BDFM
AR(1)
1.01,
1.07,
1.06,
1.00,
0.99,
0.75,
1.01,
1.02,
1.03,
1.00,
1.02,
0.65,
1.00,
1.02,
1.04,
0.97,
0.99,
0.85,
0.98,
1.04,
1.07,
0.98,
1.00,
0.47,
5, 7
9, 11
19, 16
0, 0
2, 3
0, 0
s(0)a(0)
s(1)a(0)
s(2)a(0)
s(0)a(1)
s(1)a(1)
1.01
1.02
1.03
1.00
1.00
0.75
1.01
1.02
1.02
1.00
1.01
0.53
0.97
1.01
1.03
0.96
0.98
0.78
1.01
1.02
1.02
1.01
1.01
0.40
Benchmark
0.02, 2.63
0.01, 3.17
0.07, 1.19
0.01, 2.84
4, 2
BDFM
BDFM
BDFM
BDFM
BDFM
AR(1)
0.90,
1.04,
1.08,
0.98,
0.99,
0.66,
0.91,
0.96,
0.97,
0.96,
0.99,
0.49,
0.97,
0.98,
1.00,
0.97,
0.97,
0.85,
0.92,
1.02,
1.13,
0.99,
1.00,
0.49,
1, 1
0, 2
13, 5
4, 5
0, 3
2, 2
Panel b:
2
4
s(0)a(0)
s(1)a(0)
s(2)a(0)
s(0)a(1)
s(1)a(1)
0.96
1.02
1.03
0.98
1.00
0.87
1.02
1.00
0.98
1.03
1.01
0.82
0.98
0.99
1.01
0.98
0.99
0.87
0.98
0.99
1.00
1.01
1.01
0.71
Benchmark
0.01, 3.05
0.00, 3.69
0.07, 1.30
0.01, 3.21
19, 21
BDFM
BDFM
BDFM
BDFM
BDFM
AR(1)
1.01,
1.07,
1.09,
1.00,
1.00,
0.59,
0.95,
0.98,
0.99,
0.99,
0.99,
0.45,
1.00,
1.02,
1.04,
0.96,
0.97,
0.82,
0.94,
1.03,
1.13,
1.00,
1.00,
0.37,
6, 5
2, 8
22, 17
1, 2
0, 0
0, 1
s(0)a(0)
s(1)a(0)
s(2)a(0)
s(0)a(1)
s(1)a(1)
Benchmark
0.99
1.03
1.03
0.99
1.00
0.69
0.01, 2.84
1.02
1.01
0.99
1.02
1.01
0.48
0.01, 3.45
1.00
1.01
1.03
0.96
0.98
0.78
0.07, 1.22
0.99
1.00
1.01
1.01
1.01
0.36
0.01, 3.04
8, 6
Note: Panel a) reports the results for h = 4, and Panel b) reports the results for h = 8, where h refers
to the number of lags used in equation 2. Benchmark is BDFM s(2)a(1). s() denotes the number of lags
used for the factors in the observation equation, a() denotes the number of lags used for the idiosyncratic
AR process. The abbreviations Y, E I and W are respectively GDP, employment, investment and wages
in mainland Norway. AR(1) is a univariate AR(1) models for each variable. For each model, variable,
and horizon the reported numbers are relative RMSE (left) and - average logScore (right) scores, i.e.
Mi,H,v
BDF M s(2)a(1)H,v for i = 1, · · · 6 and v = (Y, E, I, W ). For the BDFM s(2)a(1) model the numbers
reported are the actual scores. The numbers in the last column show how many times model i, at horizon
H, is ranked as the best model when the performance across all variables v = 1, · · · , N is evaluated.
34
Generally, forecasting performance increases with the number of lagged factors, while
the inclusion of autocorrelated idiosyncratic errors seems to be less important for forecasts
four quarters into the future. Viewed from a bias-variance trade-off perspective, this is
intuitive. The richer specified Benchmark model has a better in sample fit, thus a lower
bias, but may have a higher degree of variance. At longer forecasting horizons, this reduces
forecast accuracy.
Bai and Wang (2012) show in a simulation study that specifying a BDFM without
autocorrelated idiosyncratic errors, although the underlying data generating process has
this feature, generally produces estimates of the latent factors that are less reliable than
specifying a BDFM with autocorrelated idiosyncratic errors, despite the underlying data
generating process lacking this feature. Thus, although the BDFM s(2)a(0) specification
also performs well in terms of forecasting, we prefer the Benchmark model.
Evaluating the Benchmark model across Panel (a) and (b), i.e., with h = 4 and h = 8,
we see that the results are somewhat better in Panel (b). That is, the logScore is generally
higher, indicating a better density fit (while the RMSE is essentially unchanged). The
findings reported in section 5 are qualitatively similar, irrespective of whether we use h = 4
or h = 8. However, as documented in Hamilton and Herrera (2004), when modeling the oil
market, an overly restrictive lag structure might lead to misleading results. Accordingly,
we report the results for the h = 8 specification.
For many variables, e.g., GDP, simple time series models such as AR processes are
often difficult to outperform with respect to forecasting performance. We therefore also
compare the performance of the Benchmark model with that of a simple univariate AR(1)
model.22 As can be seen from Panel (a) and (b) in Table 5, the forecasting performance of
the dynamic factor model is substantially better than the AR(1). For example, at horizon
2, and for GDP in Mainland Norway (Y), the performance of the Benchmark model is over
20 and 10 percent better than the AR(1) model when evaluated using RMSE and average
logScores, respectively. For wages in Mainland Norway (W), the Benchmark model is
even more superior, with an improvement of over 50 percent relative to the AR(1) model
at horizon 4.
In summary, the results reported in table 5 support our Benchmark model specification. The highly parameterized, and structural, factor model is also superior to simple
AR(1) models for most variables and at most horizons. As such, our findings confirm
a voluminous literature documenting the usefulness of factor models for forecasting, see,
e.g., Stock and Watson (2002).
C.2
What is global activity?
As described in Section 3, we construct the observable world activity series based on the
mean across 8 different countries. These countries are not chosen ad-hoc: they represent
Norway’s most important trading partners and the largest economies in the world. That
being said, world activity is not an observable variable. Thus, we have attempted to
estimate the world activity factor as a latent factor in the same manner as we estimate
22
We estimate one AR(1) model for each observable variable, v = 1, · · · , N , and conduct the same quasireal-time forecasting experiment as described above.
35
the latent oil and domestic activity factors. This did not work well. Employing reasonable
uninformative priors, and without restrictions on the hyper-parameters, the model is not
able to distinguish the different factors from each other in any meaningful manner. Our
approach of approximating world activity as the mean across 8 different countries could
be regarded as employing more informative priors and placing restrictions on the hyperparameters. Ideally, this should have been performed within the modeling framework.
However, as the extraction of the world activity factor is not the main research question
of this study, we have not pursued the issue further.
Importantly, our main results are robust to different world activity approximations,
with one exception. China should not be excluded from the set. As shown in Aastveit et al.
(2012), growth in emerging economies (here represented by China), has been fundamental
in explaining the surge in oil prices over the last two decades. To capture this important
driver of the oil market, China should not be excluded from the construction of the global
activity factor. Including or excluding countries other than the US and China from the
international set, does not alter our main conclusions.
C.3
Subsample analysis
For production variables we have data going back until the beginning of the 1980’s. Thus,
for comparison we estimate the BDFM with production data only, i.e. excluding employment, wage and investment series, on the two samples 1986:Q1 to 2012:Q4 and 1996:Q1
to 2012:Q4. We stress that extending the sample all the way back to the 1980s is not
uncontroversial. In the 1980s the Norwegian exchange rate was more or less fixed, and the
central bank was not targeting inflation. Further, as numerous papers have documented,
both the volatility of foreign shocks and the degree of business cycle synchronization was
different in the 1980s compared to today. Thus, the comparison between the two samples
is only conducted as part of our robustness analysis. Further, the information set used to
extract the latent factors effectively becomes much smaller when employment, wage and
investment series are excluded from the analysis.23
Nevertheless, Table 6 compares the average non-mainland variance decompositions
for the two periods. One finding stand out. The domestic economy’s dependence on oil
specific and international shocks have increased over the sample. Domestic activity shocks
explain roughly 45 percent of the variation of key domestic sectors when we estimate the
model over longest sample, and only around 25 percent when we estimate the model over
the shorter sample. The results also show that world activity shocks are more important
today than in previous periods, while the variance explained by the oil activity shock is
more or less unchanged.
23
Due to the smaller information set we also use the h = 4 model specification in this exercise.
36
Table 6. Variance decompositions: Short versus long sample comparison
Shock
Sample
Sector
Oil
activity
Oil
specific
Global
demand
Non-oil
activity
1986:Q1-2012:Q4
1996:Q1-2012:Q4
Average
Average
0.27, 0.26
0.22, 0.23
0.20, 0.16
0.23, 0.16
0.14, 0.12
0.32, 0.35
0.40, 0.46
0.23, 0.25
Note: Each row-column intersection reports median variance decompositions for horizons 4 (left) and 8
(right)
As alluded to above, the results reported in Table 6 are not directly comparable to the
once reported in Table 1. Still, the results for the shorter sample are consistent with our
earlier findings, although the fraction of variance explained attributed to the different oil
specific and international shocks differ.
37
Appendix D
The Gibbs sampling approach
The three steps of the Gibbs sampler, described in Section 4.2.1, are iterated until convergence. Below we describe the three steps in more detail. The exposition follows Kim
and Nelson (1999) closely, and we refer to their book for details.
For convenience, we repeat some notation: ỹT = [y1 , · · · , yT ]0 , f˜T = [f1 , · · · , fT ]0 ,
H = [λ0 , · · · , λs ], and pi = [ρ1,i , · · · , ρl,i ] for i = 1, · · · , N , and rewrite the state space
model defined in equation 1 and 2 as:
yt = ΛFt + t
(8)
Ft = AFt−1 + et
(9)
and
0
]0 , et = Gut , with ut ∼ i.i.d.N (0, Q) and:
where Ft = [ft0 , · · · , ft−h

 
Iq
φ1 φ2 · · · φh
0
 Iq 0 · · · 0 

 

, G =  .  , Λ = H 0N,h−s
A=

.
.
 .. 
 0 Iq . . .. 
0
0 0 Iq 0

(10)
Note that h > s in our application.
We also allow for serially correlated idiosyncratic errors. In particular, we consider
the case where t,i , for i = 1, · · · , N , follows independent AR(l) processes:
t,i = pi Et,i + ωt,i
where ωt,i is the AR(l) residuals with ωt,i

σ12

0
R=
 ..
.
0
∼ i.i.d.N (0, σi2 ). I.e.:

0 ··· 0

.
σ22 . . 0 

. . . . . . ..  ,
. 
2
· · · · · · σN
(11)
(12)
and Et,i = [t−1,i , · · · , t−l,i ]0 .
D.0.1
Step 1: f˜T |ỹT , Λ, A, R, Q, p
We employ Carter and Kohn’s multimove Gibbs sampling approach (see Carter and Kohn
(1994)). Because the state space model given in equations 8 and 9 is linear and Gaussian,
the distribution of FT given ỹT and that of Ft given Ft+1 and ỹt for t = T − 1, · · · , 1 are
also Gaussian:
FT |ỹT ∼ N (FT |T , PT |T )
(13)
Ft |ỹt , Ft+1 ∼ N (Ft|t,Ft+1 , Pt|t,Ft+1 ),
38
t = T − 1, T − 2, · · · , 1
(14)
where
FT |T = E(FT |ỹT )
(15)
PT |T = Cov(FT |ỹT )
(16)
Ft|t,Ft+1 = E(Ft |ỹt , Ft+1 ) = E(Ft |Ft|t , Ft|t+1 )
(17)
Pt|t,Ft+1 = Cov(Ft |ỹt , Ft+1 ) = Cov(Ft |Ft|t , Ft|t+1 )
(18)
Given F0|0 and P0|0 , we obtain FT |T and PT |T from the last iteration of the Gaussian
Kalman filter:
Ft|t−1 = AFt−1|t−1
(19)
Pt|t−1 = APt−1|t−1 A0 + GQG0
0
0
Kt = Pt|t−1 Λ (ΛPt|t−1 Λ + R)
(20)
−1
(21)
Ft|t = Ft|t−1 + Kt (yt − ΛFt|t−1 )
(22)
Pt|t = Pt|t−1 − Kt ΛPt|t−1
(23)
I.e., at t = T , equation 22 and 23 above, together with equation 13, is used to draw
FT |T .
We draw Ft|t,Ft+1 for t = T − 1, T − 2, · · · , 1 based on 14, where Ft|t,Ft+1 and Pt|t,Ft+1
are generated from the following updating equations:
Ft|t,Ft+1 = E(Ft |Ft|t , Ft|t+1 )
0
= Ft|t + Pt|t
A(APt|t A0 + GQG0 )−1 (Ft+1 − AFt|t )
Pt|t,Ft+1 = Cov(Ft |Ft|t , Ft|t+1 )
= Pt|t + Pt|t A0 (APt|t A0 + GQG0 )APt|t
D.0.2
(24)
(25)
Step 2: A, Q|ỹT , f˜T , Λ, R, p
Conditional on f˜T , equation 9 is independent of the rest of the model, and the distribution
of A and Q are independent of the rest of the parameters of the model, as well as the
data.
By abusing notation, we put the transition equation in SUR form and define:
y = Xβ + (26)
where y = [f1 , · · · , fT ]0 , X = [X1 , · · · , XT ]0 , = [1 , · · · , T ]0 and β = [β1 , · · · , βq ]0 , with
βk = [φ1,k , · · · , φh,k ] for k = 1, · · · , q. Further,


xt,1 0 · · · 0


 0 xt,2 . . . ... 


Xt =  .
.. 
.. ..
.
.
. . 
 .
0 · · · · · · xt,q
0
0
with xt,k = [ft−1
, · · · , ft−h
]. Finally, ∼ i.i.d.N (0, Iq ⊗ Q).24
24
With the transition equation specified in SUR form it becomes easy to adjust the VAR(h) model such
that different regressors enter the q equations of the VAR(h).
39
To simulate β and Q, we employ the independent Normal-Whishart prior:
p(β, Q) = p(β)p(Q−1 )
(27)
p(β) = fN (β|β, V β )
(28)
p(Q−1 ) = fW (Q−1 |v Q , Q−1 )
(29)
where
The conditional posterior of β is:
β|y, Q−1 ∼ N (β, V β )I[s(β)]
with
V β = (V −1
β +
T
X
Xt0 Q−1 Xt )−1
(30)
(31)
t=1
and
β = V β (V −1
β β +
T
X
Xt0 Q−1 yt )
(32)
t=1
I[s(β)] is an indicator function used to denote that the roots of β lie outside the unit
circle.
The conditional posterior of Q−1 is:
−1
Q−1 |y, β ∼ W (v Q , Q )
(33)
vQ = vQ + T
(34)
with
and
Q=Q+
T
X
(yt − Xt β)(yt − Xt β)0
(35)
t=1
D.0.3
Step 3: Λ, R, p|ỹT , f˜T , A, Q
Conditional on f˜T , and given our assumption of R being diagonal, equation 8 result in N
independent regression models.
However, to take into account serially correlated idiosyncratic errors, and still employ
standard Bayesian techniques, we need to transform equation 8 slightly.
Thus, for i = 1, · · · , N , conditional on p, and with l = 1, we can rewrite equation 8
as:
∗
yt,i
= Λi Ft∗ + ωt,i
(36)
∗
with yt,i
= yt,i − p1,i yt−1,i , and Ft∗ = Ft − p1,i Ft−1 , and Λi being the i-th row of Λ.
From 36 we can then simulate the parameters Λi and Ri,i = σi2 = h1i using standard
independent Normal-Gamma priors (for notational convenience we drop the subscript i
from the expressions below):25
p(Λ, h) = p(Λ)p(h)
(37)
25
Note that with l = 0, we could have simulated the parameters Λi and σi2 without doing the transformation
of variables described above.
40
where
p(Λ) = fN (Λ|Λ, V Λ )
(38)
p(h) = fG (h|s−2 , v h )
(39)
The conditional posterior of Λ is:
Λ|ỹ, h, p ∼ N (Λ, V Λ )
(40)
with;
V Λ = (V
−1
Λ
+h
T
X
0
Ft∗ Ft∗ )−1
(41)
t=1
and
Λ = V Λ (V −1
Λ Λ+h
T
X
0
Ft∗ yt∗ )
(42)
t=1
The conditional posterior for h is:
h|ỹ, Λ, p ∼ G(v h , s−2 )
(43)
vh = vh + T
(44)
with
and
PT
− ΛFt∗ )0 (yt∗ − ΛFt∗ ) + v h s2
(45)
vh
Finally, conditional on Λ and h, the posterior of p depends upon its prior, which we
assume is a multivariate Normal, i.e.:
s=
∗
t=1 (yt
p(p) = fN (p|p, V p )
(46)
Accordingly, the conditional posterior for p is:
p|ỹ, Λ, h ∼ N (p, V p )I[s(p)]
with
V p = (V
−1
p
+h
T
X
Et0 Et )−1
(47)
(48)
t=1
and
p = V p (V
−1
p p
+h
T
X
t=1
41
Et0 t )
(49)
D.0.4
Prior specifications and initial values
The Benchmark model is estimated using two-step parameter estimates (see Section 4.2)
as priors. We label these estimates OLS. In particular, for equations 28 and 29 we set
β = β OLS , V β = VβOLS × 3, Q = QOLS and v Q = 10.
For equations 38, 39 and 46 we set v h = 10, s2 = s2,OLS , Λ = [λOLS
: 0N,h−s−1 ] and
0
Λ
p
V = [(Is × 3) ⊗ VλOLS
]. p = 0, and V = 0.5.
0
In sum, these priors are reasonable uninformative, but still proper. We have also
experimented with other prior specifications, e.g. using Minnesota style prior for the
transition equation parameters, and setting Λ = 0. This yields similar results as the once
reported in the main text. However, the variables in our sample display very different
unconditional volatilities. The prior specification should accommodate this feature.
The Gibbs sampler is initialized using parameter values derived from the two-step
estimation procedure. Parameters not derived in the two-step estimation (i.e. p and
λ1 , · · · , λs ) are set to 0.
In this model, a subtle issue arises for the t = 0 observations (i.e. lags of the dynamic
factors and the idiosyncratic errors at time t = 1). However, since we assume stationary
errors in this model, the treatment of initial conditions is of less importance. Accordingly,
we follow common practice and work with the likelihood based on data from t = h +
1, · · · , T .
42
Centre for Applied Macro - and petroleum economics (CAMp)
Centre for Applied Macro - and Petroleum economics (CAMP)
will bring together economists working on applied macroeconomic issues, with special
emphasis on petroleum economics.
BI Norwegian Business School
Centre for Applied Macro - Petroleum economics (CAMP)
N-0442 Oslo
http://www.bi.no/camp
CAMP Working Paper Series
ISSN: 1892-2198